|Publication number||US7110492 B2|
|Application number||US 10/952,800|
|Publication date||Sep 19, 2006|
|Filing date||Sep 30, 2004|
|Priority date||Oct 7, 2003|
|Also published as||DE60329105D1, EP1522959A1, EP1522959B1, US20050074090|
|Publication number||10952800, 952800, US 7110492 B2, US 7110492B2, US-B2-7110492, US7110492 B2, US7110492B2|
|Inventors||Alex Ulyanenkov, Stanislaw Sobolewski|
|Original Assignee||Bruker Axs Gmbh|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (2), Non-Patent Citations (4), Referenced by (1), Classifications (8), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims Paris Convention priority of EP 03 022 421.6 filed Oct. 7, 2003 the complete disclosure of which is hereby incorporated by reference.
The invention relates to a method of determining parameters of a sample by X-ray scattering comprising the steps of:
Such a method has been disclosed in A. Ulyanenkov, K. Omote, J. Harada; Physica B 283 (2000), 237–241.
X-ray analysis is applied in numerous ways in biology, chemistry, physics and materials science. The interaction of x-rays with a sample can reveal information about the sample that would be difficult to obtain by other means, if at all. Moreover, x-ray analysis is a non-destructive method, making it particularly attractive.
One way to investigate unknown parameters t1, . . . tN of complex samples such as multilayer structures by X-ray analysis are the so called trial and error methods. These methods start with the measurement of an experimental scattering spectrum of the sample. Then, numerous test structures (each with a different set of parameters defining a respective test structure) are assumed and simulated scattering spectra for these test structures are calculated on the basis of a physical model (e.g. kinematic scattering theory). The parameters defining the test structure whose theoretical spectrum has the best match with the experimental scattering spectrum are considered to be approximately the true parameters of the sample. The best match is typically identified as the global minimum of an error value, in particular a χ2 difference function of the experimental spectrum and the simulated spectra. Typical parameters thus approximated include film or layer thickness, lattice strain, contamination or doping level and the like.
In general, it is impossible to calculate simulated spectra for all possible test structures due to the finite capacity of computers. On the contrary, it is necessary to provide a decision algorithm determining the test structures to be calculated. There is an arbitrarily chosen test structure (or a set of test structures) to begin with, and the next test structures to be calculated are chosen by the decision algorithm on the basis of the quality of match of the previous test structures.
The most common decision algorithms are gradiental methods. If for a certain parameter ti the difference function χ2 decreases when changing the parameter ti slightly into one direction, then the next set of parameters is chosen such that ti is shifted slightly in this direction. If the difference function χ2 increases again, however, the parameter ti is no more changed into said direction. Typical gradiental methods include the Simplex algorithm, the Levenberg-Marquardt algorithm and the Newton algorithm. However, gradient methods run the risk of getting trapped in local minima not identical with the global minimum of the difference function, and no useful approximation for the true parameters of the sample can be found in such a case.
One type of decision algorithm overcoming this problem is the so called genetic algorithm. This decision algorithm works as follows:
In the beginning, there is an arbitrarily chosen starting set of so called individuals (or chromosomes). Each individual represents a test structure. These individuals all carry an equal number of so-called encoded parameters (or genes). The encoded parameters are each correlated to a physical parameter of the test structure, e.g. an impurity level or a layer thickness. Thus the number of encoded parameters corresponds to the number of physical parameters of the modeled sample to be investigated. All individuals as a whole represent a population. A typical population comprises 20 to 50 individuals.
During the genetic algorithm, the population is altered. The starting population is transformed into a new population, and this new one is altered into another one and so on. The populations at different steps of alteration are called generations. A generation to be altered is called a parent generation, and the generation newly created out of the parent generation is called an offspring generation. Thus, the starting population is the first parent generation in the algorithm. Subsequent generations of individuals are intended to contain an individual(s) approaching the global minimum of the error value.
For the creation of an offspring generation, at first the error value of each individual in the parent generation is calculated. The elite, i.e. a number of typically one or two individuals with the smallest error value in the parent generation, are copied without any alteration into the offspring generation. Then the individuals of the parent generation are subjected to genetic operators.
These operators include a selection operator which determines individuals of the parent generations which are used to produce the offspring generation.
Another operator is the mutation operator. It decides about changes in single genes of an offspring individual, in particular which genes and to what degree these genes are altered. In the state of the art, these mutations are random changes in the genes.
A third operator determines the crossover of the genes of two mating parent individuals (crossover or mating operator). It decides what gene is taken from which parent. In the state of the art, it is known to cut the parent chromosomes (=individuals) in half, and the upper part of the offspring individual is chosen as the upper part of the first parent individual, and the lower part of the offspring individual is chosen as the lower part of the second parent individual.
Through application of these genetic operators, an unlimited number of subsequent generations can be created. Typically, about 10 to 100 generations are calculated. In the final (youngest) offspring generation, the individual with the lowest error value corresponds to the test structure with the best match with the true sample structure. The genetic algorithm is very reliable and gives a good approximation to the true sample structure, with only a low risk of running into a local minimum of the error value function.
In order to increase the final accuracy of the determined sample parameters, a gradiental method may subsequently be applied, with the set of parameter determined in the genetic algorithm as a starting set of parameters in the gradiental method. This starting set of parameters should be close enough to the global minimum to avoid trapping in a non-global minimum.
However, the genetic algorithm requires a lot of calculations, determined by the amount of individuals per generation, the number of generations and the complexity of the genetic operators.
It is the object of the present invention to improve the genetic algorithm such that it can approximate the true sample parameters faster, i.e. within less generations and/or with less individuals per generation, and with better reliability.
This object is achieved by a method as described above, characterized in that a “movement” genetic operator is applied from one generation to the next to move at least some of the encoded parameters of at least some of the individuals towards respective encoded parameters of the individual with the smallest error value.
The proposed movement genetic operator is somewhat similar to the mutation operator, but the change in the gene (=encoded parameter) or genes is not random, but intentional and directed. The alteration of an encoded parameter affected by the inventive movement operator is such that the difference between the encoded parameter of the parent individual and the encoded parameter of the best matching individual of the parent generation is larger than the encoded parameter of the offspring individual and the encoded parameter of said best matching individual of the parent generation. In most cases, the inventive movement will lead to an offspring individual with a lower error value than a randomly mutated offspring individual.
The inventive movement of an encoded parameter may be set to the digital minimum, i.e. the encoded parameter is moved only one digit towards the value of the best matching individual, or it may comprise a certain percentage of the encoded parameter difference, such as 20%. Preferably, the percentage is less or equal to 50%.
It is possible to affect all or just a fraction of the individuals of a generation by the movement operator. Moreover, it is possible to alter all or only a part of the genes of an affected individual. Preferably, each affected individual is subjected to the movement operator only with respect to a few, in particular only one encoded parameter.
Note that the inventive movement operator can be applied to a parent individual immediately before creating an offspring individual, or directly at the creation of the offspring individual.
In a particularly preferred variant of the inventive method, after a given number k of successive generations the genetic operator of “mutation” is no longer applied in evolution of further generations.
In other words, in the last few generations calculated, the mutation genetic operator is deactivated. By this means, random changes in the genes are avoided when the encoded parameters and the respective test structure are already close to the corresponding desired true structure. During creation of the last generations, random changes most often worsen the quality of match of the corresponding test structure and lead away from the wanted global minimum of the error value function.
When dispensing with the random changes during the last generations in accordance with the invention, the exact global minimum can be approximated by the genetic algorithm with a higher accuracy, since only genes already close to this minimum are available for new combinations in the offspring generations. The inventive genetic algorithm during the last generations can replace a subsequent application of a gradiental method and thus save calculation power and time.
When a total number of G generations is created by the inventive genetic algorithm, and after k generations the mutation operator is no more applied, then k may be chosen such that a minimum number of generations G−k is performed without the mutation operator, for example with G−k≧5, or such that a minimum fraction of generations (G−k)/G is performed without the mutation operator, for example (G−k)/G≧20%. Alternatively or in addition, it is also possible that the transition generation (i.e. the determination of k) can be coupled to a minimum quality of match of the best fitting individual of the transition generation: as soon as the quality of match of the best fitting individual of a generation reaches a limiting value, subsequent generations are generated without the mutation operator.
In a preferred variant, at least some of the offspring individuals of every successive generation evolve from more than two parent individuals of the respective previous generation. In particular, genes from three parent individuals can be combined to generate a new offspring individual. By this means, the genes of selected individuals can be mixed in less generations, and a wider parameter space can be explored.
Further preferred is a variant of the inventive method, wherein at least some of the offspring individuals of every successive generation evolve from only one parent individual of the respective previous generation (=“cloning”). Genetic operators such as mutation and movement may still apply, however. Thus, successful gene combinations can be kept even when they do not belong to a kept elite, and can be slightly changed with a high chance of being an individual of low error value in the offspring generation again.
In an alternative and advantageous variant of the inventive method, the “crossover” genetic operator is modified such that the encoded parameters of each offspring individual is selected randomly from the respective encoded parameters of the parent individuals of the previous generation. The vector describing the random selection of the parent individual for each gene is called the random chromosome. The random choice of the genes accelerates the mixing of selected successful genes and promotes the exploration of a wide parameter space.
It is particularly advantageous for a variant if from one generation to the next at least the individual with the smallest error value is transferred unchanged (=“elite”). By this means the individual with the lowest error value in the offspring generation always has a lower or equally low error value as the individual with the lowest error value in the parent generation. Therefore the matching of the best matching individuals can only improve in the course of the generations.
A preferred variant of the inventive method provides that the error value is generated by means of a weighting function, in particular based on logarithmic values of the scattering data. By means of the logarithmic scale, an overestimation of regions of high absolute values in the error value function is avoided. The weighting function is preferably a logarithmically scaled instead of absolute x-ray intensity within the χ2 difference function.
Further advantages can be extracted from the description and the enclosed drawings. The features mentioned above and below can be used in accordance with the invention either individually or collectively in any combination. The embodiments mentioned are not to be understood as exhaustive enumeration but rather have exemplary character for description of the invention.
The invention is shown in the drawings:
In the following, the inventive extended genetic algorithm and its application to x-ray analysis are explained in detail.
The classic scheme of genetic algorithm is extended to improve the robustness and efficiency of the method. New genetic operators implemented in this work are shown to increase the convergence speed and reliability of the optimization process. A complex model function with multiple local extrema and real x-ray reflectivity and diffraction data have been used to testify the modified algorithm. The effectiveness of the new technique is compared to other optimization methods.
X-ray metrological methods in science and industry have been proved to be efficient techniques for sample characterization and growth process development and control. A large variety of sample structures can be probed by x-rays to examine film thickness, interface roughness, crystallographic lattice strain and distortion, material contamination, etc. Although the measured x-ray data can be used directly for evaluation of some sample parameters, the detailed knowledge about the sample structure can only be obtained using special data treatment procedures. These procedures usually utilize a trial-and-error technique, which uses a parameterized sample model for the simulation of the x-ray scattering process, and then the difference χ2 between calculated and measured intensities is minimized in view of the sample parameters. Thus, an effective and robust optimization algorithm is required for accurate data interpretation. Moreover, in most of experimental setups, x-ray measurements provide the magnitude of scattered x-ray intensity, whereas the amplitude and the phase of x-rays are lost, and that is why no procedures are available for reconstruction of sample physical parameters directly from measured intensities. The loss of x-ray phase information during the measurements causes the ambiguity of data interpretation results, e.g. when several sample models result in a comparably small difference χ2 between measurements and simulations. This situation requires from the optimization technique to be able to find not only a single global minimum, but also a set of deepest minima on the χ2 hypersurface, which may contain real physical solutions. Genetic algorithms (=GA) [Ref. 1], widely used nowadays, seem to be most successful in solving all mentioned problems in x-ray data analysis. They combine a stochastic search and strategy aimed features that help to find the global minimum along with other local minima of the cost function χ2 of comparable magnitude.
In the present work, we propose specific modifications of the classic GA scheme, which improve the efficiency of the algorithm. These new implementations into standard GA have been testified and the results are compared with results obtained by classic GA (CGA). As the test objects, a complex model function possessing multiple minima, and real x-ray diffraction and reflection measurements have been used. We describe and qualitatively ground the implemented modifications of the genetic algorithm. Test results of the proposed method are presented called further extended GA (XGA) for model multi-minima function and the comparison with CGA. The real x-ray diffraction and reflection measurements are fitted by using extended GA and various methods, and the performance of XGA is compared with other optimization techniques [Ref. 4].
An explanation of the details of genetic algorithms can be found in numerous publications (see, e.g. [Ref. 1] and citations therein), and in this section only the principle construction of a classic GA is described. Genetic algorithms exploit the Darwin's evolutionary idea and the principle of survival of most adapted individuals in a population. The first step in any GA application is the formalization of the parameters to be investigated in order to unify all the operations with them. Usually, the binary representation of the whole parameter space is used, that allows to easily code the problem. Every point in the parameter space then represents a unique physical state of the investigated system, and the goal is to explore this space to find the point which gives the best fit according to the pre-defined fitness criteria. Since the fitted-parameters are formalized, any set of them comprehensively describing the system is considered as an individual; the limited number of individuals compose a population, which is evolving on the basis of some genetic rules. The procedure of parameter formalization is called encoding, whereas reverse operation of obtaining the physical parameters is a decoding.
The typical GA search procedure consists of the following steps, depicted in
All of the operators described above have various implementations and internal parameters, which are used to optimize the GA efficiency for different applications. The genetic algorithm, being correctly tuned for the concrete problem, demonstrates excellent results as an optimization and search procedure [Ref. 2]. In this work, we propose further modifications of basic GA operators and general rules to increase the efficiency and reliability of the algorithm. These modifications are:
[1.] One more basic GA operator is implemented, the movement operator. Before the selection of parent individuals for mating, a limited number of parents are moved towards the best individual in the population, i.e. the parameters of the moved parents are changed closer to the parameters of the fittest individual. This operator improves the convergence of GA, by increasing the amount of individuals in the vicinity of the fittest one.
[2.] Each offspring can have not only two parents but arbitrary number of parents. In our case, we particularly propose a number of parents n=1, 2, 3. This multiple choice combines the classic two-parental approach with single “cloning” (n=1) of parents, which improves the convergence of GA by increasing an amount of individuals with better fitness, and multi-parental options (n=3), which increase the variety of a population by wider exploration of the parameter space.
[3.] A new crossover principle, providing a better mixture of parents' features in children, is implemented by using a random sequence of bits, the so called random chromosome. When parents mate, the offspring inherits the parameter from that parent, to which the corresponding bit in the random chromosome points to. This modification results in a more efficient parameter mixture in the crossover procedure, optimizes the source code and accelerates the algorithm because only simple boolean operations are involved (in particular in the case of two parents).
[4.] A more effective principle for the formation of new generations is developed. The sequence of formation is the following: firstly, the limited number of elite individuals is moved to the successive population; then the population is filled by the individuals from the previous generation and by new offsprings randomly, with a probability depending on the individual's fitness and the number of children left. This improvement prevents individuals with good fitness from being replaced by new children and replaces the individuals with rather bad fitness, i.e. increasing the convergence and decreasing the probability for the best solutions to be skipped.
[5.] The mutations are forbidden in a given number of last generations. This modification permits to use solely GA for a fine parameter fitting cycle. Because of their stochastic nature, conventional GA schemes rarely deliver exactly the point with the best fitness in the parameter space. Usually, the gradiental optimization methods have to be further applied in the vicinity of the solution found by GA to refine the parameters [Ref. 2] (see flowchart in
Thus, our modified XGA version of genetic algorithm has the following principle scheme (flowchart in
To optimize a source code, the new generation is constructed on the basis of the previous one, firstly by selection of unchangeable elite E individuals, then by substitution of old individuals by C children in dependence on the fitness of replaced individuals and the number of remained children. The evolution proceeds by repeating the cycle in
To study an effectiveness of above-mentioned XGA implementations, we used firstly a smooth two-dimensional analytical function
possessing single global maximum at (x;y)=(0.75; 0.625) and multiple local maxima (
XGA has been also applied for the fitting of real experimental data from x-ray reflectivity (XRR) and high-resolution x-ray diffraction (HRXRD), the two most commonly used x-ray techniques. The convergence and speed of XGA are compared with other known methods, simulated annealing (SA) and the simplex method (SM).
X-ray diffraction measurements (open dots) from a Si1-xGex/Si (004) Bragg reflection at wavelength λ=0.154056 nm are shown in
The inventive extended genetic algorithm is shown to be the most effective and robust optimization technique in comparison with the classic genetic algorithm, simulated annealing and the simplex method. Phenomenological tests with complex model functions possessing multiple extrema, as well as with real experimental x-ray data, both reflectivity and diffraction, have shown the advantages of XGA. In view of large computer resources required for fitting of x-ray data, the robustness and quickness of XGA plays an essential role for precise data analysis in x-ray analytical software.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US6192103 *||Jun 3, 1999||Feb 20, 2001||Bede Scientific, Inc.||Fitting of X-ray scattering data using evolutionary algorithms|
|US6823043 *||Jun 25, 2002||Nov 23, 2004||Panalytical B.V.||Determination of material parameters|
|1||*||Ghoshray, S; Yen, K.K; "More efficient genetic algorithm for solving optimization problems" 1995, IEEE Conference on Systems, Man and Cybernetics-vol. 5, pp. 4515-4520.|
|2||*||Pan Wang et al.: "Study on a novel hybrid adaptive genetic algorithm embedding conjugate gradient algorithm" Proceedings of the Third World Congress of Intelligent Control and Automation, vol. 1, Jun. 28, 2000-Jul. 2, 2000, pp. 630-633, XP010505285 China *p. 630-p. 632*.|
|3||*||Prahlada-Rao B B et al.: "A genetic algorithm for channel routing using inter-cluster mutation" Evolutionary Computation, 1994. IEEE World Congress on Computational Intelligence, Proceedings of the First IEEE Conference on Orlando, FL, USA Jun. 27-29, 1994, New York, NY, USA, IEEE, Jun. 27, 1994, pp. 97-103, XP010122866 ISBN: 0-7803-1899-4 *p. 97-p. 98* *p. 101*.|
|4||*||Ulyanenkov A. et al.: "Genetic algorithm: Refinement of X-ray reflectivity data from multilayers and thin films" 6th International Conference on Surface X-Ray and Neutron Scattering (SXNS-6) ; Noordwijkerhout, Neth Sep. 12-Sep. 17, 1999, vol. 283, No. 1, Sep. 12, 1999, pp. 237-241, XP002269236 Phys B. Condens Matter; Physica B: Condensed Matter 2000 Elsevier Science Publishers B.V., Amsterdam, Netherlands *p. 237-p. 238*.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|EP2464962B1 *||Aug 13, 2010||Feb 19, 2014||Con-Boys OY||Method and system for analysing data obtained using scattering measurements from disordered material|
|International Classification||G01N23/201, G06N3/12, G01N23/20|
|Cooperative Classification||G06N3/126, G01N23/20|
|European Classification||G01N23/20, G06N3/12G|
|Sep 30, 2004||AS||Assignment|
Owner name: BRUKER AXS GMBH, GERMANY
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:ULYANENKOV, ALEX;SOBOLESWKI, STANISLAW;REEL/FRAME:015855/0642
Effective date: 20040729
|Feb 16, 2010||FPAY||Fee payment|
Year of fee payment: 4
|Mar 13, 2014||FPAY||Fee payment|
Year of fee payment: 8